Learn How Claude Code Usage Limits Work and Make Your Usage Last Longer
Claude Code context limits need compaction; plan allowances reset by time and are shared. Check /usage, reset offers, and plan rules.
Key Takeaways
- Distinguish context limits from plan limits: a new chat or `/compact` can free context, but it will not restore a five-hour or weekly shared allowance.
- Diagnose the warning before acting: use `/context` for context issues, and `/usage` or Settings > Usage for plan resets, paid options, API billing, or reset offers.
- Know reset rules by plan: Pro and Max can have both five-hour and weekly limits, while verified Ultra accounts have a weekly limit only; special reset offers may restore allowances early without changing normal reset times.
- Read `/usage` carefully: plan bars reflect account allowance, local attribution is approximate, and recommendation percentages do not mean guaranteed savings.
- Make usage last longer by clearing unrelated work, compacting connected tasks, choosing appropriate models and effort levels, and avoiding repeated session resets during one task.
Claude Code usage limits matter for developers and teams who want a long task to survive a full workday. Anthropic’s Claude Code can pause work when a context window fills, a five-hour session limit is reached, or a weekly plan allowance runs out. Those states are not the same: a new chat can create room for context, but it cannot refill a shared allowance, and a dollar amount in /usage may be only an estimate. Documentation checked November 7, 2026; plan rules, commands, and displays can change, so check Anthropic documentation and your account.
Start With the Message You See
If a context window warning appears, inspect the conversation with /context and consider /compact, or start a fresh session for unrelated work after saving needed details. If a plan limit warning appears, open /usage or Settings > Usage to identify the affected limit and reset time; depending on the account, a separate reset or paid option may be available. If /usage shows a dollar figure, check how you authenticated and your provider’s billing. The session figure is an estimate, not necessarily a charge.
Context Window vs Plan Allowance
A context window is like a working desk: it holds instructions, messages, files, and tool results used in a conversation. A plan allowance measures usage over time across supported Claude products. API billing is a separate question.
If a conversation nears its context limit while the plan still has capacity, the problem is conversation size; compacting or starting a new session for unrelated work can help. If the weekly plan limit is exhausted while the conversation is not full, a new session will not refill it. Check whether the shorter five-hour window or the weekly window applies; the five-hour reset may arrive sooner. Depending on account and settings, you may wait, change plans, or use a paid option. Switching models does not restore an exhausted shared allowance.
Reset Rules for Pro, Max, and Ultra
For individual Pro and Max plans, Anthropic describes both a five-hour session limit and a weekly limit. The five-hour window begins with the first message and resets five hours later. The weekly limit resets at a fixed time assigned to the account. Verified Ultra accounts have a weekly limit without the five-hour limit, and other conditions can apply. Check Settings > Usage. Usage depends on model, effort, conversation length, and features used. Anthropic’s plan guide explains these plan rules.
Some eligible accounts occasionally receive a separate limit reset that can restore a specified five-hour or weekly allowance early. If available, Anthropic says it appears under Settings > Usage on the web or Claude Desktop; check the offer and any expiry before using it. It does not change the normal weekly reset time or refund charges. The reset option is not currently available inside Claude Code, though a reset used in Claude also affects the shared Claude Code allowance. See Anthropic’s limit reset guide for details and eligibility.
What /usage Actually Measures
On supported subscription plans, /usage can show plan bars, activity statistics, an attribution breakdown, and recommendations. The available details depend on the plan and Claude Code version. The plan bars represent the account allowance. The detailed attribution and recommendations are approximate and based on this machine’s local session history. Activity on another computer or on claude.ai may not appear there, even though Claude and Claude Code can share the allowance. A gap between the local breakdown and the full account total is not necessarily an error.
A percentage beside a recommendation is easy to misread. If a recommendation shows 25%, it means the recommendation applies to roughly a quarter of the recent usage covered by the analysis. It does not mean following it will save 25%. A request may appear in several recommendation rows, so percentages can total more than 100%. Attribution shares, which connect usage to features such as subagents or individual Model Context Protocol servers, answer a different question from recommendation shares. Anthropic’s cost guide explains both.
Authentication and Dollar Estimates
Authentication matters. Anthropic warns that an ANTHROPIC_API_KEY in the environment can cause Claude Code to use the API instead of a subscription, with API charges. A subscription user may want an allowance to last until its reset; an API user may be managing a bill. A dollar estimate and a plan percentage can describe the same task in different measurements. Anthropic’s plan guide explains the distinction.
Make Context, Models, and Tools Last Longer
Long conversations carry useful history—failed attempts, test output, and decisions—but unrelated work can drag that history along. If you finish debugging authentication and then move to an unrelated build configuration, /clear starts a cleaner session. Save needed details in the project first. A new session may need to rediscover context, so clearing repeatedly during one connected task can be counterproductive.
For a connected long task, /compact summarizes the conversation and can be asked to preserve current failures, decisions, and test results. A summary is not a perfect copy, so check it when exact context matters. /context shows what occupies the current window and can reveal instructions or tools you do not need, but it cannot tell you that removing an item will save a precise share of the weekly allowance. Clearing or compacting also does not reset that allowance.
Anthropic’s Claude Code guidance recommends a more capable model for harder reasoning and a smaller model for routine work. A difficult production bug, a small naming change, and a documentation lookup ask for different settings. Inspect available models with /model, and where supported, use a lower effort setting for less demanding work. Do not default to the smallest model everywhere: a lighter first attempt that misses a constraint and needs several repair rounds may cost more overall. Consider implementation, verification, and correction, and keep acceptance checks the same when comparing approaches.
Subagents can handle a noisy search or long investigation and return a concise summary, making the main conversation easier to navigate. But the subagent’s requests count toward usage. Several parallel agents may each need context and independently inspect files or call tools. Use delegation for bounded, independent tasks; moving work away from the main conversation does not make it free. A small lookup may be better done directly, while a separate review or investigation may justify extra work.
Model Context Protocol connects Claude to outside tools and data sources, such as a project tracker. Those connections and their tool results can add context and work. Claude Code documentation says tool definitions are deferred by default, so not every description is loaded upfront. /context helps show what is present, and /mcp lets you inspect configured servers. Disable an unneeded server only if safe; an essential tool can be worth its cost.
Test a Change Without Trusting the Meter Alone
The source’s scenario and 25% number are illustrative, not results from a test. If a recommendation concerns a large project instruction file, inspect the file and remove only redundant or irrelevant instructions while keeping essential project rules. Then compare a repeatable task using the original and revised instructions with the same checks. Even if the file appeared in 25% of recent local activity, the saving remains an open question.
One way to test: choose the same small bug fix from the same starting commit in two separate clean copies of a project. Write down the prompt, tests, and review criteria an acceptable patch must pass. Change one thing: use the original instruction file in one version and a carefully trimmed file in the other. Keep the model, effort setting, prompt, tools, starting code, and acceptance checks the same. Run each version in a fresh session and record the setup.
Record before and after: note relevant plan bars and reset times, and save session token information and estimated cost shown in /usage where available before clearing the session. Also record test results, time, and any human correction or follow-up needed. API users should check provider usage and billing records for actual charges. Compare completed outcomes. A smaller number is not a win if the patch fails the checks or needs significant rework. Repeat across several runs when practical and look at variation rather than declaring a saving from one attempt.
This is an experiment a reader can run, not a benchmark performed for the article. Shared Claude activity can move plan bars during the comparison, the local attribution is incomplete, and caching and ordinary run-to-run variation can affect results. Plan bars may be too coarse to isolate a small task at all. If the available data cannot distinguish the approaches, the honest result is that you do not know yet.
The practical sequence is to identify the limit first, then check the number that belongs to it. When changing workflow, judge the finished result as well as the meter. Available data can guide the decision even when it cannot account for every request or promise a specific saving. The original guidance, Claude Code Usage Limits Explained: How to Make Your Usage Last Longer, was published in Towards AI on Medium, where readers continue the conversation by highlighting and responding to the story.
Meta description: Understand Claude Code usage limits, five-hour and weekly resets, and how /usage numbers explain what to do to make your allowance last longer.
Tags: Claude Code, Anthropic, usage limits, AI coding tools
Featured image: Abstract friendly illustration generated with AI, inspired by an author-created conceptual diagram based on Anthropic’s usage guidance and cost documentation, with no people, logos, or text.
Comments (0)